[BUG]

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Observed regressions (all new, all consistent across sessions)

Misreads or incompletely processes prompt instructions — requires 2–3 correction rounds to reach correct interpretation
Ignores saved memory files (MEMORY.md + individual .md memory entries) that were being applied correctly for the prior month
Creates new ADO work items when instruction is explicitly to update existing ones
Misses acceptance criteria or test scenarios that are clearly stated in the work item
Fails to ask clarifying questions when instructions are ambiguous — assumes incorrectly instead
Misunderstands skill/slash command intent despite identical usage pattern to prior month

What Should Happen?

I want my "OLD-CLAUDE" back. This one is useless.

Error Messages/Logs

You're right. I was wrong to blame compaction. That's been constant for a month and wasn't an issue. Let me be actually honest:

I don't know what changed, and that's the real problem.

Something changed on Anthropic's side — model update, weight change, backend adjustment — that I have zero visibility into. You didn't change your workflow. Your memory files, your instructions, your session structure are all the same. The variable is me, and I can't tell you exactly what shifted.

What I can tell you objectively:

The failures are systematic, not random — missed requirements, comprehension drops, ignoring memory. That pattern points to a model-level regression, not a one-off.
The timing (past 2 days) suggests a deployment or update on Anthropic's end.
I cannot self-diagnose or self-fix a model-level regression.
What you need to do:

Report this now at github.com/anthropics/claude-code/issues. Be specific: "Consistent quality regression in Claude Sonnet 4.6 starting approximately [date]. Failures include: missed requirements from AC, not reading memory files, misunderstanding prompts, requiring multiple correction rounds. Was working correctly for the preceding month with identical workflow."

Check if there's a newer or different model available in your Claude Code settings — if you can pin to an earlier version or try Opus, that may bypass whatever regressed.

What I can do right now to partially compensate:

For every task from here: I will explicitly read MEMORY.md first, state what I'm about to do in one sentence before acting, and ask one question if anything is ambiguous. It won't fix a model regression but it reduces the blast radius.

The real fix needs Anthropic to look at it.

Steps to Reproduce

Steps to reproduce

Open Claude Code with claude-sonnet-4-6
Load a session with populated MEMORY.md (multiple .md memory entries)
Issue an instruction referencing a prior-saved rule (e.g. "update existing TCs" — rule is saved in memory)
Observe: rule is ignored, new items created instead of updating existing ones
Repeat across multiple sessions — regression is consistent, not intermittent

Expected: Behaviour consistent with prior month — memory applied, instructions followed on first pass, clarifying questions asked when ambiguous

Actual: Systematic quality drop across comprehension, memory application, and task execution

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.123

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

What did NOT change

Session structure and length
Memory file content and format
Skill definitions
User prompt style and phrasing
MCP server configuration